Abstract:

Accurately assessing borrower default status is a fundamental aspect of personal credit risk evaluation and a critical factor in lending decisions made by financial institutions. However， conventional statistical and machine learning techniques often have limited generalization capability and unstable individual model performance. Furthermore， ensemble learning methods often require labor-intensive hyperparameter configuration， considerable manual intervention and a tendency to overfit. To address these challenges， an automated， highly accurate framework for predicting personal credit default is proposed. The core methodology leverages the Auto Gluon （AGT） framework， an automated machine learning system specifically recognized for its efficacy with tabular data. The model uses a multi-layer stacking ensemble architecture， where each layer incorporates predictions from the previous layer and the original feature set. Predictions from various base models are amalgamated via a weighted aggregation mechanism， thereby automating the processes of model selection and hyperparameter optimization. Furthermore， to mitigate the prevalent issue of class imbalance in financial datasets， the framework incorporates a class-balanced modification of the cross-entropy loss function. This recalibrates the loss function in proportion to the inverse class frequency， enhancing the model’s sensitivity to minority-class instances. To evaluate the proposed AGT-multi-layer-stacking-MCE model， both proprietary and public datasets are utilized. This included real-world lending data from Lending Club and Paipaidai， as well as publicly available credit data from the UCI repository： specifically， the German， Japanese， and Australian datasets. Performance is assessed using key metrics including Type II error， the area under the ROC curve （AUC） and accuracy. Comparative benchmarks include traditional machine learning models， Super Learner ensemble techniques and previously published results on the same datasets. The empirical results show that the proposed model produced Type II errors of 0.0676 and 0.0226 on the two main datasets， representing reductions of 9.2% and 28.7% respectively compared to the most effective baseline model. The model outperformes conventional data-balancing techniques across all evaluated datasets， highlighting its robustness and superior discriminative capacity. The substantial potential of automated machine learning frameworks supplemented by tailored loss functions in advancing credit default prediction is illustrated. The AGT-multi-layer-stacking-MCE model elevates predictive accuracy， particularly in minimizing Type II error， which is critically important in financial risk assessment， and diminishes reliance on expert-driven manual tuning. Consequently， the proposed approach provides a scalable， efficient and highly adaptive solution for credit scoring， facilitating more reliable and automated decision-making in financial practices. It contributes to the broader adoption of advanced machine learning techniques in operational environments， offering meaningful insights into how to handle class imbalance in financial risk modelling.

Key words: default prediction, online lending, AutoGluon, stacking ensemble